Why Enterprise Agentic AI Pilots Stall and How Enterprises Move Them to Production
Enterprise agent programs stall when an agent has to perform reliably across a claims queue, a loan pipeline, or an anti-money-laundering (AML) alert backlog. Building the agent is no longer the hard part, since managed platforms and open-source frameworks supply models, tool calling (letting an agent act in other software), and orchestration (coordinating the steps of a task) from the first day.
Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs or unclear business value, or because teams lack adequate risk controls (Gartner, 2025).
In McKinsey’s 2025 survey, 39% of respondents attribute any level of earnings before interest and taxes (EBIT, a measure of operating profit) impact to AI, and most of those say only a small share of their organization’s EBIT is attributable to AI use (McKinsey, 2025).
Context is the missing piece. Agents built on standard operating procedures (SOPs, written step-by-step instructions) and interviews, supplemented by event logs (timestamped records of system activity), often miss exception paths, spreadsheet workarounds, and the cross-application handoffs that define how regulated work runs.
This post covers insight from our whitepaper, “The Strategic Role of First-Party Data in Enterprise Agentic”; download the full whitepaper now.
What is enterprise agentic AI?
An enterprise agentic AI platform is the software layer an organization uses to build, run, orchestrate, and govern autonomous, goal-directed AI agents in production, combining a reasoning model, orchestration, memory, governed data access, connectors, observability, and guardrails. Forrester defines agentic AI as systems of foundation models, rules, architectures, and tools that “flexibly plan and adapt to resolve goals by taking action in their environment, with increasing levels of autonomy” (Forrester, 2025).
How is agentic AI different from LLMs, RPA, and traditional AI/ML?
Agentic AI plans and acts toward a goal, whereas a general-purpose large language model (LLM) generates a response, robotic process automation (RPA) replays scripted steps, and traditional machine learning (ML) predicts an outcome. Gartner divides the work by type: “using AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval” (Gartner, 2025).
| Approach | What it does | How it handles exceptions | Where it fits |
|---|---|---|---|
| General-purpose LLM | Generates text or answers from a prompt, with no tool access in a direct call | Can produce plausible but incorrect output | Summarizes or drafts content. Answers questions. |
| RPA | Executes if-then-else steps on structured data through the UI | Stops, or escalates to a person | Stable, high-volume steps that follow rules |
| Traditional AI/ML | Predicts or classifies from historical data | Underperforms when conditions shift away from the training data | Scores or forecasts outcomes. Flags fraud. |
| Agentic AI | Plans multistep work and calls tools to take actions | Reasons through case variants within guardrails and escalates at defined thresholds | Reviews claims or loan files with mixed data. Also reviews alerts. |
Gartner also warns about “agent washing,” which it describes as rebranding assistants, RPA, and chatbots “without substantial agentic capabilities” (Gartner, 2025). Test whether each product plans and calls tools. Check whether it completes actions on its own.
What architecture does an enterprise agentic AI platform need?
An enterprise agentic AI platform typically has six layers:
- Reasoning: A foundation model interprets the goal and plans steps. It chooses which tool to call.
- Orchestration: A runtime sequences steps and routes tasks between agents. It holds state for long-running work.
- Memory: Session state carries the active case, and long-term memory keeps facts across sessions.
- Data access and RAG: Retrieval grounds the model in enterprise records through permissioned connectors.
- Observability: Traces and logs show what the agent did. Evaluation scores show how well it performed.
- Guardrails: Content filters and tool-call policy checks limit what the agent can do. Per-agent identity constrains its access.
Orchestration coordinates agent and human work, while memory and governed context supply the case information agents need for execution.
Orchestration and multi-agent coordination
The orchestration runtime assigns automated steps to an agent or a deterministic workflow (a fixed, scripted sequence of steps) and routes human steps to a person, and because a claim or loan file can stay open for days, it has to persist case state (keep a record of where the case stands) between steps. Forrester describes the governing layer as a control plane (the layer that oversees and directs agents) that “inventories, governs, orchestrates, and assures heterogeneous AI agents across vendors and domains” (Forrester, 2025). In regulated operations that plane also has to cover human queues, because every escalated case ends up with a person.
Memory, data foundation, and governed context
Memory and retrieval are preconditions for any agent, and both are limited to what someone has written down. Formal documentation often leaves out the tacit knowledge operators use to handle exceptions. An agent that retrieves the SOP will still miss the step a senior adjuster takes when a policy record and a repair estimate disagree.
One way to fill that gap is a continuously updated model of observed work (described in the platform section below). Machine-usable operating procedures derived from it encode steps, decision rules, exception handling, and escalation thresholds for the agent to execute.
How do agents connect to enterprise systems?
Agents reach systems that expose application programming interfaces (APIs), which let one program call another, through standard protocols, and they reach legacy systems through the screen.
The Model Context Protocol (MCP) connects agents to tools and applications, with access to their data (Linux Foundation, 2025), and enterprise resource planning (ERP) and customer relationship management (CRM) systems with APIs connect through MCP servers or existing connectors. The Agent2Agent (A2A) protocol handles communication between agents; the Linux Foundation identifies version 1.0 as its first stable specification (Linux Foundation, 2026).
Where systems lack usable APIs, teams can evaluate screen-level access. Computer-use agents interpret screenshots and request mouse or keyboard actions, which the application executes (OpenAI, 2025). For workflows spanning mainframes (large legacy central computers) and sessions on virtual desktop infrastructure (VDI, remotely hosted desktops), desktop agents using computer vision can document cross-application handoffs. They can also capture spreadsheet workarounds and exception paths before teams authorize agent execution.
How do these platforms hold up to governance in regulated industries?
These platforms pass governance review when each level of agent autonomy maps to a control an auditor can test:
- Audit trails: The platform records every input, output, tool call, decision and rationale at the case level.
- Role-based access: Each agent has its own identity and least-privilege permissions scoped to its task.
- Escalation and override: Defined thresholds route cases to a named person, who can override or reverse its decisions. That person can also halt the agent.
- Runtime guardrails: Policy checks run at the tool-call boundary, outside the agent’s reasoning, before any action executes.
- Governed autonomy: Autonomy expands only as evaluated performance earns it, and humans keep the judgment calls.
The same controls map to the frameworks examiners already use:
| Agent control | SOC 2 Type 2 (2017 TSC) | HIPAA | ISO/IEC 27001:2022 Annex A | EU AI Act (high-risk) |
|---|---|---|---|---|
| Audit trails | CC7.2 anomaly monitoring | §164.308(a)(1)(ii)(D) information system activity review | 8.15 Logging; 8.16 Monitoring activities | Art. 12 automatic logs; Art. 26(6) log retention of at least six months |
| Role-based access | CC6.7 restriction of information transmission, movement and removal to authorized users and processes | No single standard in this map | 5.15 Access control; 5.18 Access rights; 8.2 Privileged access rights | Art. 26 oversight assigned to competent natural persons |
| Escalation and override | CC7.3 security event evaluation and response | §164.308(a)(1)(ii)(D) review of security incident tracking reports | 5.25 Assessment and decision on information security events | Art. 14(4) authority to override, reverse or halt through a “stop” button |
The table maps three of the five controls. Use the map as a starting point. NIST cautions that crosswalks “are not always one-to-one” (NIST, 2020). Observation tools built on privacy by design redact sensitive screen fields inside the customer environment, before data leaves it, which keeps captured process data within the customer perimeter.
The Commission’s AI Act Service Desk states that the rules for AI systems and general-purpose AI models “also apply to AI agents” (European Commission). Credit scoring and life and health insurance pricing are Annex III high-risk uses whose obligations apply from December 2, 2027 (EUR-Lex, 2026).
Why do so many enterprise agent pilots fail to reach production?
Agent pilots stall when they move from a controlled demo to live cases, because live cases surface work no one documented.
- Undocumented exceptions: The SOP covers the standard path, and the agent meets a case variant no one wrote down, such as a claim with conflicting coverage dates.
- No machine-readable ground truth and context decay: Documentation is stale and interviews do not scale. Event logs record only what instrumented systems capture, so the team cannot specify or benchmark correct behavior. Processes also drift after launch, and an agent built on a one-time snapshot keeps executing yesterday’s procedure until outcomes degrade.
These gaps trace to the gap named at the start of this post: the agent has no record of how the work runs outside the documentation.
How do enterprises move agents from pilot to production?
Enterprises move agents to production through a staged observation-to-agent pipeline with an exit criterion at each gate. Few organizations formalize those gates. Only 31% of CIOs report having guidelines for deciding when an AI pilot should not move to production (Gartner, 2026).
- Observe: Capture how the target process runs across every application. Exit when the team has documented variants and exception paths, along with a human performance baseline.
- Prioritize: Score each task for rules-based automation versus judgment. Exit when the team has a ranked list tied to cost and cycle-time impact.
- Encode: Convert observed work into machine-usable operating procedures. Exit when each procedure covers the observed variants, decision rules, exception paths and escalation thresholds, and a process owner signs off.
- Simulate and benchmark: Run the agent on historical cases and compare its task success and escalation rates with the human baseline. Compare error rates as well. Exit when the agent meets the agreed threshold.
- Launch under control: Promote the agent with scoped permissions and human review of escalations. Keep full audit logging. Exit when live metrics hold against the baseline.
- Watch for drift: Keep comparing live execution with the observed process. Re-baseline and return to Simulate whenever live execution deviates from the observed process.
These six steps can be mapped to three key maturity stages:
- Explore covers Observe, Prioritize, and Encode, produces first results in weeks, and ends when a process owner signs off on the procedures.
- Expand runs Simulate and benchmark and Launch under control for the first workflows, with promotion when live metrics hold against the human baseline.
- Enterprise Scale extends the pipeline to further workflows, each promoted on the same baseline test, with Watch for drift running continuously.
What does it cost to run agents at scale?
Agents cost more per task than chat because iterative reasoning and multi-step tool use multiply token consumption. Gartner projects that routing a task to an agentic reasoning model raises provider inference costs at least five times compared with a basic chatbot interaction (Gartner, 2026).
A total cost of ownership (TCO) model should cover:
- Licensing: Platform subscriptions.
- Token consumption: Planning, retrieval, tool calls, reasoning and guardrail checks.
- Runtime infrastructure: Persistent compute and memory stores, including session state.
- Observability and evaluation: Log ingestion and trace retention, plus evaluation runs.
- Human review: Staff time for escalations and quality sampling.
Vendor-hosted software as a service (SaaS) carries less operating burden, with data residency (the requirement that data stay in a given location) depending on the provider’s hosting options. The private-cloud deployment model keeps traffic on private networks, and the fully customer-hosted deployment model supports data residency requirements at the cost of up-front investment and full operational responsibility.
Which platforms and frameworks do buyers evaluate, and where does Skan AI fit?
Buyers compare managed cloud agent platforms, low-code agent builders, process orchestration suites, and open-source frameworks against the same criteria, including a source of operational context beyond documents and event logs. Process orchestration suites add agent coordination, and open-source frameworks suit teams building their own runtime. Whatever the category, test each option against these criteria:
- Native MCP and A2A support
- Per-agent identity, least-privilege permissions, audit logs, human-in-the-loop thresholds, a kill switch and rollback
- Built-in evaluation against defined success criteria
- Token cost tracking per workflow
- A source of operational context beyond documents and event logs
Skan AI is an agentic AI platform: it builds a model of how work runs across people, applications, and decisions. Here’s how it works:
- The platform combines end-to-end process visibility with zero-integration capture to to observe process activity across mainframes, VDI, legacy desktop applications, web applications, and spreadsheets.
- Skan Virtual Assistant (Skan VA) uses desktop-level computer vision (software that reads what appears on a screen), with privacy by design.
- From that observation, and without backend connectors, Skan AI builds the Context Graph of Work: a continuously updated model connecting people, applications, tasks, decisions, policies, outcomes, and governance constraints.
- Agentic operating procedures (AOPs) derived from that observed behavior give agents steps and decision rules to execute. They also define escalation thresholds.
In contrast, process mining reconstructs workflows from system records, reaches end-to-end depth only after integrating with every system of record (the system holding the authoritative copy of the data), and it misses work that never reaches an event log; task mining only watches individual user tasks, captures only those individual tasks.
The platform has three key products, all powered by the Context Graph of Work
- Skan AI Blueprint maps activity and ranks AI and automation opportunities.
- Skan Intelligence covers end-to-end flows and variant analysis. It detects bottlenecks and drift.
- Skan AI Agents executes approved work against the Context Graph of Work and AOPs. Agent Studio defines scope and permissions, along with escalation paths.
The platform is built for end-to-end privacy by design. Skan PRISM, an on-premises gateway, applies pixel-level redaction inside the customer environment as part of privacy by design, so raw screen data stays inside the customer perimeter, and Skan AI does not capture email content or keystrokes. Enterprises can choose on-premises inside the customer firewall, private cloud, or Skan AI hosted.
Which workflows deliver measured value first?
The first workflows to show measured value combine high volume, mixed structured and unstructured inputs, defined policy, and a measurable human baseline:
- Claims adjudication: Agents assemble the file and validate coverage. They route complex or disputed cases to adjusters. The industry baseline is slow: U.S. homeowners wait more than a month before final payment on property claims (J.D. Power, 2026).
- Loan origination: Agents gather documents and check conditions. They prepare the file for the underwriter’s decision.
- AML/KYC alert review: Agents collect evidence across systems and draft dispositions for analyst approval.
- IT operations: Agents triage tickets and run approved runbook steps.
- Compliance monitoring: Agents flag deviations from approved procedures.
Production evidence should include task success and escalation rates measured against the human baseline captured before launch. Compare mean time to resolution against that baseline as well.
See the pipeline applied to your own process
Bring a claims or loan workflow and see how the observation-to-agent pipeline applies to it. An alert-review workflow can also serve as the starting point. Request a Skan AI demo or download the full whitepaper.
Frequently Asked Questions
No single platform leads every category. Gartner tracks AI agent development and agent management platforms as separate markets. Choose based on your workflows, deployment constraints, governance requirements, and access to operational context beyond documents and event logs.
Enterprises build agents with managed cloud agent platforms, low-code agent builders, process orchestration suites and open-source frameworks, and buyers often mix built and bought components. Compare them on MCP and A2A support, per-agent identity and audit logs, built-in evaluation and per-workflow token cost tracking, using your own workflow data.
Yes, Google’s documentation describes Gemini Enterprise Agent Platform as “an evolution of Vertex AI” (Google Cloud, 2026), and Google announced it on April 22, 2026 (Google Cloud, 2026).
Start with high-volume workflows that mix rules with judgment and have a measurable human baseline, such as claims adjudication or loan origination. AML/KYC alert review is another candidate. Prioritize processes where the agent writes back to a system of record. Promote an agent only after it matches the human baseline on task success and escalation rate.
